I've been seeing Claude pop up a lot in conversations about bitcoin development and tech stuff. So, I'm curious for the folks who are actually building, reviewing code, or involved in Bitcoin projects, how's it working out for you?
Does it really assist with things like getting a grip on bitcoin core, looking over BIPs, debugging, or doing some security research? Or is it more of a tool for explanations and exploration instead of serious dev tasks?
I really wanna know how you guys are weaving it into your usual workflow.
Never use AI to operate wallets or to handle pull requests / issues on sensitive codebases such as Core.
Linux kernel is a bit different since the worst that can happen is crashes, but on Bitcoin Core, the code is money. So this is why such projects should stick with mainly human reviewers while using AI to help understand the prognosis and assist in creating bugfixes only.
I used it to draft this testnet softfork proposal and to write the code for it. I also periodically use it to write custom scripts that allow me to calculate useful info, such as for observing the sustainability of the network overtime.
Yes, it's capable of undertaking pretty serious tasks, there just needs to be someone responsible to prompt it right, and review the output.
I think that using closed hosted LLMs is not very compatible with the ethos of Bitcoin. In addition the the direct effect of any work done historically any work w/ bitcoin has prepared you to do *future* work as necessary. But if you work using a closed hosted LLM you only get the direct benefit and potentially get worse at actually doing the work-- you become dependent.
You also allow unknown third parties realtime remote access to your systems and the ability to effectively impersonate you.
If you're using it on a system with your wallet(s) vs development in a sandbox a hosted LLM absolutely blows your privacy because the providers will log and retain all information that goes in and out. I know from first hand experience that commercial LLM providers retain data far beyond what you would expect from their public statements and will provide it in response to absolutely anyone carrying a subpoena.
I would suggest as an alternative using Qwen-3.6-27B-- many regard it as competitive with Opus 4 and it has open weights and can be run on a workstation class gpu. Even if you currently don't have the resources to run it locally using a hosted copy at least improves the autonomy since you can run locally if push comes to shove and you have many choices of where to host.
I can say first hand that Qwen running with a vm-sandboxed hermes is capable of very sophisticated development work.
a >= 24 gb consumer class gpu (AMD 7900XTX, Nvidia 4090/5090) will run qwen 3.6-27b q4 locally too. but the context window may not be enough for really big tasks if youre used to workstation stuff.
Claude can only be useful to check the code and find some hidden bugs and flaws in the code.
Someone already used this tool to exploit some shitcoins, and I am sure hackers are using it steal coins from exchanges.
But totally replacing any human bitcoin developer with Claude or any other AI would be stupid and dangerous.